MINIMAX H3 / TESTED AUGUST 24, 2026
Face and costume held.
Motion timing shifted.
This MiniMax H3 Web example combines an anime image with a five-second motion reference. Inspect the inputs, output and six scoring criteria to see what survived and what changed.
This score describes one generated video. Repeatability and output quality under the same conditions in Runway or Seedance 2.5 are untested. The example was generated on the Web, not locally.
Fixed inputs
One original anime-style still
The asymmetric jacket, hair ornament, bag, leg straps and differently colored shoes make changes easier to spot.
FILE: benchmark-reference.png
One five-second motion clip
An original mannequin sequence with arm raises, horizontal extensions, a one-leg lift and side steps, filmed by a fixed camera. Five seconds, 24 fps, portrait 9:16, no audio.
FILE: benchmark-motion-reference.mp4Generation conditions
- Test date
- August 24, 2026 (JST)
- Environment
- MiniMax Web / Omni Reference / MiniMax H3
- Requested output
- 5 seconds / portrait 9:16 / 2K
- Actual output
- 5.17 seconds / 1440×2560 / watermarked
- Time and credits
- About 2 minutes / 120 free-tier credits consumed
- Seed and build
- No fixed seed or internal build number was available in the Web UI. Exact reproduction is unverified.
- Input image SHA-256
8baa48a7071c6dcb0e6f3d00a05b7362bd37288e141abd0ad50409d587dcc084- Motion clip SHA-256
2c5e212f8c6d275b2328313e052bd4eb760a8f155d2da9931e8ef9c49eaec0dc- Result SHA-256
d1a75981d6e7f3f59585d035cffc56654473955ccd50e55111b1f840d4259af4
These measurements belong to this one Web generation, not a local GPU test or a current price estimate. The hashes identify the exact input and output files.
On the same date we loaded the references in Runway with Seedance 2.5 selected, but the available free plan did not allow generation. The approximately 18-second input analysis is not a video-generation time; no credits were consumed. Runway and Seedance were one access check, not two independent generation tests.
Scoring rubric
| Criterion | What is checked | Points |
|---|---|---|
| Face and style | Contours, eyes, hair and rendering remain consistent with the input image | 20 |
| Costume and props | Clothing shapes, patterns, accessories and props do not disappear or mutate | 15 |
| Hands and anatomy | Fingers and limbs have plausible counts, connections and positioning | 20 |
| Motion fidelity | Pose order, timing and weight shifts follow the reference video | 20 |
| Camera | Framing follows the instruction without unintended zoom, rotation or angle changes | 10 |
| Temporal stability | Flicker, background deformation and character substitutions are limited | 15 |
Preliminary result: 84 / 100
Face, hair, clothing and background stayed consistent in the sampled frames. Arm raises, extensions and a one-leg lift appeared, but the order and timing of the motion peaks differed from the reference.
| Criterion | Score | Observation |
|---|---|---|
| Face and style | 19 / 20 | Contours, hairstyle, eyes and rendering remained consistent across the ten sampled frames. |
| Costume and props | 14 / 15 | The asymmetric jacket, bag, leg straps and differently colored shoes were preserved. |
| Hands and anatomy | 17 / 20 | No obvious extra limbs or major disconnections. Hands approached the frame edge during the horizontal arm extension. |
| Motion fidelity | 11 / 20 | Arm raises, horizontal extensions and a one-leg lift appeared, but their order and peak timing differed from the reference. |
| Camera | 9 / 10 | The fixed camera and portrait 9:16 framing were preserved without unintended zoom or rotation. |
| Temporal stability | 14 / 15 | The background and character stayed stable without conspicuous substitutions or major flicker. |
| Total | 84 / 100 | Preliminary score for one example, not a tool ranking. |
What this result cannot tell you
- Whether different artwork or motion will succeed consistently.
- Which tool has better output quality, success rate or cost per usable result.
- Whether local H3-Base can match the Web version's resolution or output.
CHOOSE YOUR WORKFLOW
Compare the inputs and environment you need.
Browser access, reference support and output length can narrow the shortlist without making unsupported quality claims.